Datasets / Fall Detection Dataset
Fall Detection Dataset v1.2 Real falls have reflexes.Your training data should too.
Synthetic fall detection dataset, with a commercial licence and no real people in the data.
Free sample and datasheet. Licence warranty: 30 days after delivery.
What is the Fall Detection Dataset?
The Xentim Fall Detection Dataset is a synthetic dataset for training skeleton-based fall detectors and person-down alarms. It holds 12,108 sequences and 2.22 million annotated frames, with 17 COCO keypoints, a bounding box and a fall phase for each person on each frame, from 12 virtual cameras per event.
People brace when they fall.
Simulated falls include the reflex of a hand reaching for the ground, so your model sees more than a body dropping.
Trained to ignore what is not a fall.
Non-fall activity makes up 46.6% of sequences, so normal movement can stay quiet.
Ready for your pose estimator.
Keypoints follow COCO 17 at 1920x1080, for pipelines that run on RTMPose or ViTPose.
How it compares with public fall datasets
| Xentim Fall Detection | CAUCAFall | GMDCSA-24 | |
|---|---|---|---|
| Data | Synthetic skeletons | Real video | Real video |
| Licence | Paid commercial licence | CC BY 4.0 | MIT on GitHub |
| Best use | Training at scale, 12 cameras per event | Testing on real footage | Testing on real footage |
Sources: CAUCAFall, GMDCSA-24. Checked 30 October 2026. Check each licence yourself before use.
What you get
| Simulations | 1,009 events, 12 virtual cameras each |
|---|---|
| Annotations | 17 COCO keypoints, visibility, bounding box, fall phase, 3D skeleton |
| Splits | train 9,972, validation 2,208, test 1,332 sequences, split by base motion |
| Files | index.csv, frames.parquet, one JSON per sequence, report |
| Not included | RGB images or video, this is skeleton data |
What it is for
Simulated falls include the reflex of a hand reaching for the ground, so your model sees more than a body dropping.
Choose your licence.
Same dataset, different rights. Licences are perpetual.
Standard
One legal entity, all its sites. Updates v1.x for 12 months. Email support.
Order StandardEnterprise
Everything in Standard, plus use by your affiliates, priority support and one custom scenario pack of up to 200 simulations, delivered within 6 weeks.
Order EnterprisePrices exclude VAT. Sold to companies and institutions. Read the full licence or open the datasheet. Questions? contact@xentim.com
Does it include video?
No. The dataset holds skeleton data, with no RGB images or video. It suits models that classify keypoints from a pose estimator.
Will it work on real footage?
Validate on your own footage before deployment. The datasheet recommends adding detector-like noise during training, and we cannot promise a set accuracy.
Can synthetic data train a fall detector?
Yes, as long as you validate on real footage. A skeleton-based classifier learns from keypoints, so it can train on simulated motion and then run on keypoints from a pose estimator such as RTMPose.
Does it include sideways falls and near misses?
Yes. The dataset includes sideways collapses, pushes, near misses and everyday non-fall activity.
Which keypoint format does it use?
COCO 17 keypoints in pixel coordinates of a 1920x1080 image, with a visibility flag, plus a 3D skeleton.
Can I test it before buying?
Yes. Each dataset has a free sample and a datasheet, so you can check the format and labels on your own model first.
How do I order?
Choose a licence and send us your order. We reply with an order form and invoice, you pay by bank transfer, SEPA or card, and we send a secure download link within 3 business days of payment.
What if the data does not match the datasheet?
Tell us within 30 days of delivery. We correct the dataset or, if we cannot, refund the fee against deletion of all copies.
Other datasets
Fights
1,045 events, 10 kinds of fight and 10 look-alikes such as hugs and high fives.from 3,000 EURSee the datasetRail
1,142 events across 3 station scenes, with falls and pushes onto the track.from 3,000 EURSee the datasetFire and smoke
89 video clips, 10,822 annotated frames, with steam and fog labelled as look-alikes.from 2,900 EURSee the datasetLast updated 30 October 2026
Test it on your model first.
The sample is free. If the data does not match its datasheet, you have 30 days to tell us.